MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning | Ayoung Lee et al. | ResearchPod

Transcript: MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

Alex: Welcome to another episode of ResearchPod. Today, we're looking at a paper on a deceptively tricky problem: why does a moral choice that seems perfectly reasonable in one country feel completely wrong in another?

Sam: It's a challenge that cuts to the heart of how AI systems are built. The paper argues that current AI models are culturally tone-deaf—they rely on a single, largely Western way of thinking about right and wrong. The authors propose a new method to help AI shift its "moral compass" to match the cultural context of the person it's talking to.

Alex: So this is fundamentally about stopping AI from quietly imposing one set of values on the whole world?

Sam: That's exactly it. Right now, most systems treat moral reasoning as a one-size-fits-all task. When you translate a moral dilemma into another language, you often just get the words—the cultural context, the customs, the social expectations, all of that gets stripped away. The authors argue that to make sound judgments, an AI needs to reason within the user's own cultural framework, not just translate into it.

Alex: That makes sense. If you're asking an AI for guidance on a sensitive situation, you'd want it to understand the unwritten rules of that situation, not just the literal meaning of your words.

Sam: Right. To address this, the researchers built a new benchmark called MCLASH. Think of it as a large library of moral dilemmas—but instead of simply translating them from English, they culturally adapted each one. If a scenario references a specific company or a local custom, they swap it for something that actually makes sense in that culture. It's not a translation; it's a cultural reconstruction.

Alex: Okay, so they have a richer, more culturally honest dataset. But how does the model actually use that information to reason differently?

Sam: That's where the core method comes in. They developed a two-step approach called MET—Multilingual Ethics with Theory-grounded reasoning. Here's the key idea: instead of just asking the model "what's the right thing to do?", they first ask it to consult a library of moral frameworks. These are established ways of thinking about ethics that philosophers and psychologists have developed over centuries—things like Confucian ethics, which emphasizes duty to family and social harmony, or deontology, which focuses on following universal rules regardless of consequences.

Alex: So before the model even tries to answer, it picks the right philosophical "lens" for that culture?

Sam: Exactly. It's like picking the right tool before you start the work. Once it selects the relevant framework, it uses that to guide its reasoning step by step. But the researchers found that just prompting the model to do this wasn't enough—the models would often take shortcuts and skip the careful reasoning. So they added a second stage called MET-D, which is a form of self-training.

MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

Ayoung Lee, Ryan Kwon, Yunxiang Zhang, Yuxuan Liu, Peter Railton, Lu Wang

Jul 13, 2026·5:47·arxiv·Original Paper

MCLASHMET (Multilingual Ethics with Theory-grounded reasoning)MET-D (MET-Distillation)Culturally Aligned Moral ReasoningMoral Theory GroundingCultural Adaptation

Listen

6 min

Briefing

Research Question

As language models are increasingly deployed in high-stakes moral decision-making, they often struggle to account for the diverse cultural and linguistic contexts of their users. Existing benchmarks rely on direct translations that strip away cultural nuance, and current reasoning methods lack grounding in formal moral theory. This paper addresses these gaps by developing a framework for culture-aware, theory-grounded moral reasoning that functions natively in multiple languages.

Approach

The researchers introduce three primary contributions:

MCLASH: A new multilingual benchmark consisting of 1,852 long-form moral scenarios adapted for six languages (Chinese, Hindi, Korean, Malay, Spanish, and English) to capture culturally situated norms.
MET (Multilingual Ethics with Theory-grounded reasoning): A two-step prompting method where the model first selects relevant moral grounds (e.g., deontology, Confucian values) from an expert-curated list and then generates a reasoning chain in the user's native language.
MET-D (MET-Distillation): A self-distillation training stage that uses a synthetic dataset of character-driven dilemmas to teach models to substantively engage with the selected moral grounds, eliminating the need for expensive human supervision.

Main Findings

MET-D significantly improves performance across all tested model families (Qwen3 and Gemma3). On the MCLASH benchmark, MET-D achieved an average macro-F1 gain of 3.71 points, with a notable peak improvement of 12.94 points for Malay. Furthermore, the method increased the use of native-language reasoning by an average of 62.13 percentage points, making the model's decision-making process more transparent and legible to non-English speaking users. The study also demonstrates that the most effective moral grounds for reasoning vary systematically across different cultures.

Why It Matters

This work provides a scalable path for aligning AI moral reasoning with local cultural values. By moving away from English-centric, "one-size-fits-all" moral frameworks, the authors demonstrate that models can be made more reliable and culturally sensitive in sensitive domains like law and healthcare. The use of self-distillation also offers a practical solution for improving reasoning in low-resource languages where human-annotated data is scarce.

Key Terms and Definitions

MCLASH — A multilingual moral decision-making benchmark consisting of culturally adapted scenarios rather than direct translations.
Theory-Grounded Reasoning — A method of decision-making where the model explicitly selects and applies established philosophical or psychological frameworks to justify its conclusions.
MET-D — A self-distillation training technique that uses synthetic, character-based dilemmas to force models to engage deeply with selected moral grounds.
Normative Authority — A dimension of moral reasoning used in this paper to establish the foundational premises or sources of legitimacy for a moral judgment.
Rejection Sampling — A training technique used here to retain only those model-generated reasoning chains that match the ground-truth decision associated with a specific character's value profile.

Abstract

Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation, failing to adapt culture-specific items; 2) inference-time methods for moral reasoning rely on static, English-centric scaffolds and lack grounding in moral theory; 3) training methods for moral decision-making typically require expensive supervision from stronger models or human annotators. We address these gaps with three contributions. First, we introduce MCLASH, a multilingual moral decision-making benchmark to capture culturally situated moral intuitions and social norms across languages. Second, we propose MET (Multilingual Ethics with Theory-grounded reasoning), a two-step prompting method built on expert-curated, theory-based grounds drawn from psychology and philosophy: the model first selects situation- and culture-specific grounds, then reasons over them in the native language of the user. Third, we introduce MET-D (MET-Distillation), which enhances the second step through a self-distillation training stage that requires no external supervision. MET-D improves macro-F1 over the base model on all three models of different sizes and families (Qwen3-4B, Qwen3-8B, Gemma3-4B), by an average of 3.71 points on MCLASH and 4.23 on MMoralExceptQA, with a peak MCLASH gain of 12.94 points for Malay on Qwen3-8B. We further reveal that MET-D increases native-language reasoning by 62.13 points on average, and that beneficial grounds differ systematically across cultures. Together, these contributions open the path for culture-aligned, theory-grounded multilingual moral reasoning.

MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning: Research Paper Podcast